Decoupled.

TL;DR this week:

The old model was simple. You ranked on Google, you got cited everywhere else. That model is dead. The new Ahrefs data on AI Overviews shows that more than 60% of pages cited inside Google's own AI answer no longer rank in the top 10 — and almost a third don't crack the top 100 at all. Gemini 3 went global in January as the AI Overview default, and citation behavior changed underneath the SEO industry's feet.

Meta's pushing the same wedge on the paid side. Their AI agent is now writing ads, picking audiences, and managing budgets while marketers approve from the sidelines. Gartner says we'll double the share of marketing work AI runs by 2028.

The throughline this week is in What The Bots Are Reading, the information gain teardown in Feed the Engine, and the 90-minute fan-out sprint in Strategy You Can Steal. AI search no longer trusts your ranking position. It's looking for something else.

Let's build.

— ZC

What's the last keyword where your top-3 Google ranking did not show up in ChatGPT or AI Overviews? Best answers go in next week's issue.

This week's term: Query Fan-Out

Query fan-out is the technique AI search engines use to split a single user question into many related sub-queries, then aggregate citations from the SERPs of those sub-queries instead of the original one.

When you ask Google AI Mode "what is the best CRM for early-stage startups," the system silently runs ten or twelve related searches behind the scenes — best free CRMs, CRMs with PLG features, CRMs that integrate with Slack, CRMs for under 10 seats — and the page that gets cited in the final answer is the one that shows up most often across that whole spread.

This is why ranking #1 for one keyword doesn't guarantee a citation anymore. The page Google cites is the page that wins the cluster of sub-queries, not the head term. It is also why the Ahrefs study this week found that 62% of AI Overview citations now come from pages that don't even rank in Google's top 10 for the original query. Those pages are winning sub-queries the user never typed.

For content marketers, the practical move is to map fan-outs for your money queries and add passages on each sub-query to your top-ranking pages. The unit of optimization is no longer the page-keyword pair. It is the page-cluster pair.

Next time someone in your Slack says "we ranked #1 but we're not in the AI Overview," you can say "your page is losing the fan-out — it's winning the head term and getting outranked on the sub-queries."

If Google is splitting your keyword into ten sub-queries and pulling citations from the SERPs of those sub-queries, you need to know what those sub-queries are and where you stand on each one. This week's sprint maps your fan-out exposure and finds the gaps where you're invisible.

Choose one money keyword where you currently rank in the top 5 on Google but do not appear in ChatGPT, Perplexity, or Google AI Overviews for the same question. Most teams have at least three of these. Write the query at the top of a doc. Write your ranking position next to it. Write "AI: 0" next to it. This is your baseline.

Open ChatGPT, Perplexity, and Google AI Mode in three tabs. Run the same query in each. For each platform, write down which URLs were cited. Then ask the AI: "What sub-questions did you research to answer this?" The platforms will usually tell you — and those sub-questions are your fan-out map. You should end this block with a list of 8 to 12 sub-queries and a list of the URLs being cited for each.

For each sub-query, pull the top three cited URLs and skim the passage being quoted. Pay attention to the entities named (products, people, concepts, places), the specific stats cited, and the structural pattern (FAQ, definition block, comparison table). Mark every entity or stat your competitor has that you don't. This is your gap list. The page that gets cited has all of these entities; your page has half of them.

Open the page that ranks #1 on Google for your target query — the page that is being skipped by AI search. Add a FAQ block with the 8 to 12 sub-queries from Block 2 as questions. Each answer should be 40 to 60 words, lead with a direct factual statement, and include at least two of the entities or stats you flagged as gaps. Republish. This is the single highest-leverage change you can make this week without rewriting the whole page.

Why this works: AI Overviews and ChatGPT cite passages, not pages. A page that has the answer to ten sub-queries written as discrete passages gets cited ten different ways. The page that has the same information buried in flowing prose gets cited zero ways. When we did this audit on our own ranking pages at Averi, the pages we added fan-out FAQs to picked up ChatGPT and Perplexity citations within three weeks, even when their Google rank didn't move. Entity density and passage structure are the new ranking factors. The old ones are noise on top.

Google has a patent called "Information Gain Score for Search Results." It is the most-cited document in the GEO research community right now, and it is the cleanest single explanation for why the Ahrefs study this week looks the way it does.

Information gain scores how much new information a page adds on top of what a model has already seen across other top-ranking pages. The page that ranks #1 on Google often has the lowest information gain in the cluster, because its content overlaps heavily with positions #2 through #10. A page ranking #14 that introduces a new stat, a new framework, or a novel data point can have higher information gain than every page above it. AI Overviews and ChatGPT both optimize for non-redundant evidence when they assemble a citation list, which is exactly why 62% of AI Overview citations now come from pages outside the top 10.

The mechanical impact splits in two directions. On the SEO side, information gain has become a quiet input to Google's Helpful Content classifier — pages that look like rewrites of competing top-10 pages get demoted in core updates. On the GEO side, information gain is closer to a primary ranking factor: the model would rather cite one page with one unique stat than five pages that all repeat the same consensus answer. Pages that pass an information gain check tend to be the ones with proprietary data, original frameworks, named expert quotes, or specific case study numbers.

One thing to do this week: Pull your three most-trafficked posts. For each one, ask what is the single most-quotable original fact on this page that does not appear on any of the top-10 ranking pages for the same query? If you cannot find one, add one. A customer outcome, a survey result, a benchmark you ran internally, a counter-intuitive observation backed by your own data. One original fact per page is the floor.

Next week: llms.txt and robots.txt — what to allow, what to block, and the new file every site needs in the root directory.

The number is 12.

Brandi AI's 2026 study of B2B publishing velocity found brands shipping 12 new or optimized pieces in a 90-day window saw up to 200x faster AI visibility gains than brands shipping four. Twelve is the threshold where citation surface starts to compound.

That number breaks every assumption SEO trained marketers to hold.

The old answer was 50 posts to start ranking and 100+ to compete on competitive terms. Twelve sounds suspiciously low until you look at what changed.

AI engines don't care about topical breadth across 100 keywords. They care about structural extractability across one cluster. A site with 12 well-architected pieces in one topic cluster looks more authoritative to ChatGPT than a site with 100 keyword-targeted blog posts spread across unrelated topics.

This shifts the entire content math for seed-stage startups. You no longer need 18 months and a content team to compete. You need 90 days and a sprint plan.

The page that ranks first is no longer the page that gets cited.

Ahrefs just dropped a study of 863,000 keywords and 4 million AI Overview URLs. The headline number is the one at the top of this newsletter: 38% of pages cited inside Google's AI Overview also rank in the top 10 for that same query. Seven months ago, that number was 76%. The other 62% of citations now come from pages that either rank deeper than position 10 or do not rank at all in the top 100.

Two things changed underneath this data. First, Google replaced its AI Overview model with Gemini 3 on January 27, 2026. Second, Google has now confirmed that AI Overviews use a "query fan-out" pattern, splitting one user query into many related sub-queries and pulling citations from the SERPs of those sub-queries instead of the main one. A page that ranks position 14 for the original query might rank position 2 for a sub-query the user never typed.

The implication is rough on traditional SEO. The page that wins the keyword no longer automatically wins the answer. The page that wins the answer is the one that shows up across the spread of related sub-queries — the one with the broadest entity coverage, the cleanest passage structure, and the highest factual density.

Other platforms have been operating this way for longer. A SearchAtlas study of 18,000 queries found ChatGPT shares only 8% of its cited URLs with Google's top 10, Gemini shares 6%, and Perplexity shares 28%. The gap is widest where AI systems have the most freedom to do their own retrieval, and the only platform that still mirrors Google rankings closely is Google's own AI Overview — and that mirror is shattering in real time.

What this means for startups: Stop optimizing for one keyword and one page. The unit of AI search optimization is the entity cluster, not the keyword page. If you publish three deep pages on related entities with rich passage structure, you can win 30 sub-queries and show up in the AI Overview without ranking number one on the main query.

Our own data at Averi backs this up — the pages we publish that get cited in ChatGPT and Perplexity are not the ones ranking #1 in Google. They are the ones with the highest entity density and the cleanest answer blocks.

📊 By the numbers

We all know the industry loves to pretend it’s got it all figured out.

But here?

We tell the truth.

Every week, we’re asking you to share how things really work behind the scenes. Because getting better starts with getting honest.

Don’t worry, it’s a safe space.

Drop your vote.

Microsoft says only 16% of brands track AI search performance. So let's find out where you actually stand.

The gap you click on is the one you're losing weekly.

We'll publish the breakdown next week and show you which Averi workflow plugs that exact hole.

Or skip the headache → Fix your workflow gaps

KitKat × VML — "The Great KitKat Heist"

On March 26, a truck carrying 413,793 KitKat bars vanished between a factory in central Italy and a distribution center in Poland. The cargo was not generic stock. The shipment was a special run of KitKat bars shaped like Formula 1 race cars, made to celebrate the 75th anniversary of F1 and the 90th anniversary of KitKat. Twelve tonnes of branded chocolate, gone.

A normal brand sends a press release expressing concern and lets the news cycle move on. KitKat and VML did the opposite. By April 1, the brand had launched a purpose-built batch-code tracker on its website. Anyone holding a KitKat could enter the eight-digit code from the wrapper and instantly find out whether their bar was part of the stolen shipment. The site doubled as a recovery effort and as the best piece of earned media the brand has run in years. KitKat US then dropped a faux movie poster called "The KitKat Job" in partnership with Tubi, riffing on Ocean's Eleven. Then a video of Secret Service-style guards protecting a KitKat truck rolling through Toronto. British Airways, Domino's, and Ryanair joined the meme thread.

Why we love it: The campaign did not invent a problem to manufacture buzz. It took a real supply-chain crisis and turned it into an interactive product moment — the wrapper became the campaign surface. The strategic insight was that consumers always had the asset (a KitKat bar with a batch code) and only needed a reason to look at it. The brand didn't ask for attention. It gave you a reason to check the back of a chocolate bar you already owned.

Results have rolled in through April: mainstream news coverage in 40+ countries, peer-brand replies from British Airways, Domino's, and Ryanair, a documented uplift in KitKat search interest, and a brand sentiment score above the brand's 24-month average per VML's reporting.

The algorithm could never: No optimizer would have approved a campaign that admits 12 tonnes of inventory is missing. No template would have suggested a batch-code lookup tool as the response surface. The whole campaign required a human to read the situation, see the opportunity in the embarrassment, and trust that consumers would meet the brand with humor instead of outrage. That trust is the one thing a model trained on best-practice playbooks cannot underwrite.

The startup takeaway: When something goes wrong publicly, the highest-leverage move is the one that turns your customers into participants in fixing it.

→ See the campaign and tracker on Famous Campaigns

FEATURED:

This is the one to be at this week. 12,500+ founders, execs, and VCs converging on the San Mateo County Event Center for three days of B2B + AI programming. For DFTA readers, the most relevant tracks are the GTM-with-AI sessions, the agentic ops tracks, and the content + GEO sessions from the marketing leaders headlining this year. If you are a founder running marketing yourself, this is the highest concentration of peers you can find in one room. The hallway track is the conference.

Date: May 12-14, 2026

Location: San Mateo County Event Center

Link: Register

The pattern this week is loud: AEO and GEO are no longer being bolted onto SEO roles. They are getting their own titles, their own budgets, and their own comp bands.

Stripe's $236K ceiling on the AEO/GEO Marketing Manager role is the highest we've seen for this title to date, and ADT is opening a director-level seat with the word "Discoverability" instead of "SEO" in the title. The job market is repricing for AI search faster than most marketing teams are repricing their content strategy.

Getting certified first → Averi Academy

Know of a GEO or AI content role that should be here? Reply with the link and we'll feature it next week.

Did You Know? —  Perplexity has 28% URL overlap with Google's top 10. ChatGPT has 8%. Gemini has 6%. The three AI search engines your customers are using do not agree with Google — or with each other — on what the best answer looks like. Optimizing for one is no longer optimizing for all.

Til next time,